All the random variables below are defined on the same probabiluty space $(\Omega, \mathcal{F}, \mathbb{P})$.

Consider the following model

$$ Y\equiv 1\{\epsilon > \beta_0+\beta_1X\} $$ where

  • $1$ is an indicator function equal to $1$ if the condition inside is satisfied and zero otherwise

  • $Y$ is a random variable with support $\mathcal{Y}\equiv \{0,1\}$

  • $X$ is a random variable with support $\mathcal{X}\equiv \{0,1\}$ (treatment)

  • $\epsilon$ is a random variable with support $\mathbb{R}$

  • $\beta_0, \beta_1$ are scalar parameters

A source I am considering states that the average treatment effect (ATE) is

$$ \begin{aligned} \text{ATE} & = P_{\epsilon}(\epsilon > \beta_0+\beta_1)- P_{\epsilon}(\epsilon > \beta_0) \end{aligned} $$ without specifying which probability is $P_{\epsilon}$.

Question: Is the probability used to compute the ATE, $P_{\epsilon}$, conditional on $X$ or unconditional?

I am tempted to say that it should be conditional; in other words

$$ \begin{aligned} \text{ATE} & \equiv E(Y|X=1)-E(Y|X=0)= \mathbb{P}(\epsilon > \beta_0+\beta_1|X=1)- \mathbb{P}(\epsilon > \beta_0|X=0) \end{aligned} $$

However, then the book claims that

$$ \begin{aligned} \text{ATE} & = P_{\epsilon}(\epsilon > \beta_0+\beta_1)- P_{\epsilon}(\epsilon > \beta_0) \\ &= \sum_{x\in \{0,1\}} \mathbb{P}(\epsilon > \beta_0+\beta_1|X=x) \mathbb{P}(X=x)\\ &-\sum_{x\in \{0,1\}} \mathbb{P}(\epsilon > \beta_0|X=x) \mathbb{P}(X=x) \end{aligned} $$ which seems to be an application of he law of iterated expectations making sense only if $P_{\epsilon}$ is not conditional on $X$.

Any hint?

  • $\begingroup$ From which book is this? $\endgroup$ Commented Jul 25, 2018 at 8:52

1 Answer 1


The ATE is defined as $ATE = \mathbb{P}(Y = 1|do(X =1)) - \mathbb{P}(Y =1 |do(X=0))$. When you intervene on $X$, you force it to be the value you want. Thus,
$$\mathbb{P}(Y= 1|do(X =1)) = \mathbb{P}(1(\epsilon>\beta_0 + \beta_1(1)) = 1) = \mathbb{P}(\epsilon>\beta_0 + \beta_1)$$

and, analogously,

$$\mathbb{P}(Y=1|do(X=0))= \mathbb{P}(1(\epsilon>\beta_0 + \beta_1(0)) = 1) = \mathbb{P}(\epsilon > \beta_0)$$


$$ATE = \mathbb{P}(\epsilon>\beta_0 + \beta_1) - \mathbb{P}(\epsilon > \beta_0)$$

Your confusion stems from defining ATE in terms of the difference in conditional expectations, an observational quantity. ATE is a structural/interventional quantity, not an observational quantity. They are only equal under certain conditions, such as $\epsilon \perp X$. The task of finding an observational expression that is equivalent to an interventional expression is what we call identification.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.